rmaf (version 3.0.1)

ma.filter: Refined Moving Average Filter

Description

uses refined moving average filter to estimate the trend component, and then obtain seasonal component if necessary.

Usage

ma.filter(x, q = NULL, seasonal = FALSE, period = NULL, plot = TRUE)

Arguments

x

a numeric vector or univariate time series.

seasonal

a logical value indicating to estimate the seasonal component. Only valid for seasonal = TRUE. The default is FALSE.

q

specified moving average lag $q$. The default is NULL.

period

seasonal period. The default is NULL.

plot

a logical value indicating to make the plots. The default is TRUE.

Value

• A matrix containing the following columns:
• data
  original data x.
• trend
  fitted trend.
• season
  seasonal indices if seasonal = TRUE.
• residual
  irregularity or residuals.

Details

For univariate time seties $x[t]$, the additive seasonal model is assumed to be $$x[t] = m[t] + S[t] + R[t],$$ where $m[t], S[t], R[t]$ are trend, seasonal and irregular components, respectively. Once we obtain the optimal moving average lag $q$ using qn, the trend can be estimated by using the refined moving average $$mhat[t] = \sum x[t]/(2q+1),$$ for $q + 1 \le t \le n - q$. If $q + 1 > n - q$, we take $q = min(n - q, q)$. If there is no seasonal component, the irregularity or residuals can be computed by $Rhat[t] = x[t] - mhat[t]$. Otherwise, the seasonal index $Shat[t]$ can be estimated by averaging the sequence $x[t] - mhat[t]$ for each of 1:period. For example, the seasonal component in January can be estimated by the average of all of the observations made in January after removing the trend component. To ensure the identifiability of $m[t]$ and $S[t]$, we have to assume $$S[i + j*period] = S[i], \sum S[i] = 0,$$ where $i = 1,..., period; j = floor(n/period)$. The irregularity or residuals are then computed by $Rhat[t] = x[t] - mhat[t] - Shat[t]$. For $t < q + 1$ and $t > n - q$, the corresponding estimators are based on equation (7) in D. Qiu et al. (2013). More details about estimating the trend component can be seen in Section 1.5 of P.J. Brockwell et al. (1991) or Chapter 6 of J. Fan et al. (2003).

For the multiplicative seasonal model $$x[t] = m[t] * S[t] * R[t],$$ it can be transformed to an additive seasonal model by taking a logarithm on both sides if $x[t] > 0$, i.e., $$log(x[t]) = log(m[t]) + log(S[t]) + log(R[t]),$$ and then use the refined moving average filter for the components decomposition as the same in the additive seasonal model.

Plots of original data v.s fitted data, fitted trend, seasonal indices (if seasonal = TRUE) and residuals will be drawn if plot = TRUE.

References

D. Qiu, Q. Shao, and L. Yang (2013), Efficient inference for autoregressive coeficient in the presence of trend. Journal of Multivariate Analysis 114, 40-53.

J. Fan and Q. Yao, Nonlinear Time Series: Nonparametric and Parametric Methods, first ed., Springer, New York, 2003.

P.J. Brockwell, R.A. Davis, Time Series: Theory and Methods, second ed., Springer, New York, 1991.

See Also

ss.filter

Examples

## decompose the trend for the first difference of annual global air temperature from 1880-1985
data(globtemp)
decomp1 <- ma.filter(globtemp)

## decompose the trend and seasonality for CO2 data with monthly and additive seasonality
decomp2 <- ma.filter(co2, seasonal = TRUE, period = 12)

## decompose the trend and seasonality for monthly airline passenger numbers from 1949-1960
decomp3 <- ma.filter(AirPassengers, seasonal = TRUE, period = 12)

## simulation data: oracally efficient estimation for AR(p) coefficients
d <- 12
n <- d*100
x <- (1:n)/n
y <- 1 + 2*x + 0.3*x^2 + sin(pi*x/6) + arima.sim(n = n,list(ar = 0.2), sd = 1)
fit <- ma.filter(y,seasonal = TRUE,period = 12,plot = FALSE)
ar(fit[,4], aic = FALSE, order.max = 1)$ar